memory: the snapshot bound ladder is inconsistent, and the CLI duplicates the byte cap

[0thernet]

Follow-up to #37. With the join no longer quadratic, work is the last limit to bind rather than the first, which exposes the other bounds. Measured while scaling algal memory query over real curated facts.

The ladder, in the order it actually binds

#	bound	value	where
1	CLI file load	262,144 B	main.rs, two call sites
2	snapshot canonical bytes	262,144 B	memory.rs query()
3	facts list	2,048	memory.rs query()
4	result rows	256	Limits::max_rows
5	join bindings	4,096	Limits::max_bindings
6	work	250,000	Limits::max_work

Two problems.

The CLI duplicates the byte cap. load(&snapshot, 262_144) in main.rs gates the file before query() ever sees it, so raising the library's limit alone changes nothing — the CLI rejects with document bytes first. Two constants govern one thing.

The facts-list bound is unreachable. At the ~170 canonical bytes a typical fact occupies, 2,048 facts is roughly 348 KB, well past the 262,144-byte snapshot cap. Bound 2 always fires before bound 3, so list(&snapshot["facts"], 2048) never actually limits anything.

Measurements

At the current ceiling, 1,429 real facts and 260,549 bytes: 167 ms, ~16 MB peak RSS, 4,628 work, which is 1% of the work budget.

Raising bounds 1 through 4 in a throwaway worktree, then scaling by relabelled copies of the same real structure:

facts	bytes	ms	work	% work	result
1,533	280 KB	—	4,992	1%	138 rows
3,066	565 KB	376	9,984	3%	276 rows
6,132	1.1 MB	753	19,968	7%	552 rows
12,264	2.3 MB	1,469	39,936	15%	1,104 rows
24,528	4.6 MB	—	—	—	Datalog join bindings (bound 5)

Work doubles exactly per doubling of facts, confirming the post-#37 linearity. At 12,264 facts the engine is at 15% of the work budget and roughly 1.5 s.

What I am not proposing

Specific numbers. These are resource-posture decisions and AGENTS.md is explicit that every count and byte size gets a bound, so raising them is yours to choose rather than mine. Two things do look like plain defects regardless of the numbers chosen:

1. the byte cap living in two places, so the library's constant is not authoritative
2. a facts-list bound that cannot be reached

Happy to send a PR for either, or for a coherent ladder once you pick the targets.

Found while building algal-bio, which uses algal memory over aging-biology curation.

🤖 Generated with Claude Code